Avalanche Search and Rescue (SaR) operations are highly time-sensitive, since the likelihood of survival rapidly declines after burial. Current localisation relies on avalanche transceivers (ARTVA) operating at 457 kHz, which offer reliable short-range accuracy but suffer from limited coverage. In this paper, we explore the use of LoRa technology as a long-range complementary approach within the LoRa-SNOW project. We introduce LS-TBLA, a CNN model that leverages temporal patterns in LoRa signal measurements, such as RSS and SNR, collected along the rescuer’s path to infer the direction of the buried victim. By encoding these measurements into statistical feature representations, the model enhances resilience to signal fluctuations and environmental noise. The approach is validated on a real-world dataset acquired in 2025, consisting of multiple search trajectories. Performance is assessed under both balanced and unbalanced spatial configurations, capturing different search conditions. The results show that exploiting temporal information leads to improved directional predictions, with consistent behaviour even in more challenging scenarios.
Fast Avalanche Search and Rescue Using LoRa Signal Dynamics and Deep Learning
Annalisa Maggini;Alexander Kocian;Stefano Chessa;Michele Girolami
2026-01-01
Abstract
Avalanche Search and Rescue (SaR) operations are highly time-sensitive, since the likelihood of survival rapidly declines after burial. Current localisation relies on avalanche transceivers (ARTVA) operating at 457 kHz, which offer reliable short-range accuracy but suffer from limited coverage. In this paper, we explore the use of LoRa technology as a long-range complementary approach within the LoRa-SNOW project. We introduce LS-TBLA, a CNN model that leverages temporal patterns in LoRa signal measurements, such as RSS and SNR, collected along the rescuer’s path to infer the direction of the buried victim. By encoding these measurements into statistical feature representations, the model enhances resilience to signal fluctuations and environmental noise. The approach is validated on a real-world dataset acquired in 2025, consisting of multiple search trajectories. Performance is assessed under both balanced and unbalanced spatial configurations, capturing different search conditions. The results show that exploiting temporal information leads to improved directional predictions, with consistent behaviour even in more challenging scenarios.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
